Miguel G. Silva, LASIGE’s PhD student, published a paper “Cutting through the noise: Explaining residuals in multivariate time series with motif analysis”, in Pattern Recognition Journal, a top-ranked journal in Pattern Recognition and Signal Processing. The paper co-authors are LASIGE’s integrated researcher Sara C. Madeira from Faculdade de Ciências, Universidade de Lisboa, and Rui Henriques from INESC-ID and IST, Universidade de Lisboa. 

The research presents a novel methodology for extracting actionable insights from the irregular components of multivariate time series, challenging conventional interpretations of residuals. Aimed at explaining irregularities in complex systems, the approach extends motif discovery to detect patterns that are statistically significant, non-trivial, and actionable. It introduces new actionable statistics to reduce search biases, rank pattern relevance, assess statistical significance under fair null models, enable robust hyperparameter tuning, explore complex multivariate residual structures, and incorporate domain knowledge via specialized masks. The method’s effectiveness is demonstrated across real-world case studies, including human activity, household energy usage, and demographic trends in Lisbon, Portugal. It highlights periods with high noise-to-signal ratios, offering plausible explanations for observed irregularities. Key findings include activity-specific motifs in human behavior, appliance usage trends, and urban mobility patterns relevant to city planning. Notably, the method uncovered non-cyclic but statistically significant motifs explaining up to 50% of the residuals variation, revealing structured patterns valuable for decision-making.

The paper is available here.